How do JSON metadata files align with the audio data?

[ErasureCode]

Hi! Thanks for the excellent work. I understand that the JSON files in dataset-annotation-json/ serve as metadata for the dataset. I was wondering:

1. What are the main fields included in each JSON entry?
2. How exactly do these metadata entries correspond to the audio files?

[xid32]

Thanks for taking the time to explore the dataset.

1. Main Fields in Each JSON Entry

Each JSON entry in the dataset contains the following fields:

• user_content:
  A text prompt describing a reasoning task. This includes a major premise, minor premise, a conclusion to evaluate, and instructions for how to decide whether the conclusion is entailed or not.

• answer:
  A natural language explanation that walks through the reasoning process and ends with the predicted label (either 'entailed' or 'not-entailed').

• label:
  A numerical label indicating the final classification:

  • 1 = entailed
  • 0 = not-entailed

• id:
  A unique identifier for the example.

2. How Do the Metadata Entries Correspond to Audio Files?

The id field in each JSON entry corresponds directly to a folder in the audio_dataset/{train|dev|test}/ directory.

Each such folder (e.g., 0001/) contains:

• question.wav: the spoken version of the user prompt, including the major premise, minor premise, and conclusion.
• answer_0.wav: the spoken version of the full chain-of-thought reasoning and final answer.

This setup enables end-to-end evaluation of audio-language models on logical reasoning tasks using both spoken input and spoken output.